Recorded synapses gave a 0.025% error decrease beyond a digital twin in a calibration test of 91,647 MICrONS V1 pairs

Can recorded synapses improve prediction of measured visual response similarity after a digital twin gives that similarity? We tested release v1 data from one MICrONS mouse. The panel had 91,647 V1 pairs, 36 source axons, 7,546 targets and 2,840 recorded connections. Both calibration rules received model movie correlation, geometry, reliability, layer and imaging scan measurements. The added rule also received connection presence, synapse count and mean cleft size. Each test source was absent from calibration training in both cell roles. Five inner source groups selected separate control and synapse penalties, with exact omission available. Synapse terms decreased squared test error by 0.02489%, from 0.02126970 to 0.02126440. A 500-draw source resample with outer refitting gave a central 95% range of -0.03412% to +0.07481%. This range kept the original penalty choices. The gain on recorded connections alone was 0.759%, with a fixed-fit source range of -0.456% to +1.792%. Without the twin term in either rule, the all-pair gain was 0.149%. Planned checks gave small gains or error increases. The tested synapse terms thus gave little mean added prediction after the given functional model and other measurements. This does not show that anatomy has no useful functional information. One mouse, selected axons, noisy responses and different video sets limit the result. The package preserves code, the complete compact input panel, source hashes, all predictions, fits, resample draws, figures and repeat instructions. Independent solvers and endpoint audits passed.

connectome · structure and function · MICrONS · digital twin · visual cortex · prediction · synapse count · synapse size · calcium imaging